Reproduce “LiquidCP: Decoupling Attention Tasks from Communication and Kernel Boundaries in Context Parallelism” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongyu Chen, Yikang Ruan, Yuhai Long, Ziyu Huang, Yiyu Luo, Liukun Yu, Yangjie Zhou, Zihan Liu and 9 more
LiquidCP schedules transfers, attention, and reductions as tasks that proceed when inputs are ready. MegaRing coalesces ready KV tiles for long queries, while MegaDCP plans history splits around dependent reductions for short-query batches. On eight H100 GPUs, core-attention throughput geometric-mean speedups over the fastest baselines were 1.23× forward and 1.18× backward for training, and 1.20× decode-only and 1.35× mixed batches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.